- GraalVM for JDK 23 (Latest)
- GraalVM for JDK 24 (Early Access)
- GraalVM for JDK 21
- GraalVM for JDK 17
- Archives
- Dev Build
- Truffle Language Implementation Framework
- Truffle Branches Instrumentation
- Dynamic Object Model
- Static Object Model
- Host Optimization for Interpreter Code
- Truffle Approach to Function Inlining
- Profiling Truffle Interpreters
- Truffle Interop 2.0
- Language Implementations
- Implementing a New Language with Truffle
- Truffle Language Safepoint Tutorial
- Truffle Native Function Interface
- Optimizing Truffle Interpreters
- Options
- On-Stack Replacement
- Truffle Strings Guide
- Specialization Histogram
- Testing DSL Specializations
- Polyglot API Based TCK
- Truffle Approach to the Compilation Queue
- Truffle Library Guide
- Truffle AOT Overview
- Truffle AOT Compilation
- Auxiliary Engine Caching
- Truffle Language Safepoint Tutorial
- Monomorphization
- Splitting Algorithm
- Monomorphization Use Cases
- Reporting Polymorphic Specializations to Runtime
Monomorphization Use Cases
This guide demonstrates through examples how monomorphization can improve performance of dynamic languages without going into any detail on how monomorphization is implemented (described in the Splitting guide) or how to leverage monomorphization in your language implementation (described in the Reporting Polymorphism guide).
Monomorphization #
To better illustrate the benefits of monomorphization, consider a small example written in JavaScript:
function add(arg1, arg2) {
return arg1 + arg2;
}
function callsAdd() {
add(1, 2);
add("foo", "bar");
}
var i = 0;
while (i < 1000) {
callsAdd();
i++;
}
As you can see in this example, the add
function is called from callsAdd
once with integer arguments and once with string arguments.
Once add
is executed enough times to be compiled its execution profile will show that the +
operator has been executed with both integers and strings and thus handlers (i.e., type checks and execution) for both types will be compiled which has a
performance impact.
This can be avoided by rewriting the example as follows:
function addInt(arg1, arg2) {
return arg1 + arg2;
}
function addString(arg1, arg2) {
return arg1 + arg2;
}
function callsAdd() {
addInt(1, 2);
addString("foo", "bar");
}
i = 0;
while (i < 1000) {
callsAdd();
i++;
}
In this example the add
has been duplicated (split) in such a way that each type profile is contained in a separate copy of the function (addInt
and addString
).
The result is that, come compilation time, only a single type profile is available for each function avoiding potentially costly type checks in the compiled code.
Automating the detection suitable candidates, as well as their duplication, performed at run time is what we call monomorphization. It is, in other words, automated run-time monomorphization of polymorphic nodes through AST duplication.
Example 1 - Monomorphization of Arguments #
This example is an extended version of the illustration example from the previous section.
The add
function is still the target for monomorphization and is called from the action
function 3 times with 3 sets of different arguments (numbers, strings and arrays).
Execute the action
function for 15 seconds in order to have enough time for warmup, and afterwards execute it for 60 seconds keeping track of how long each execution took, reporting finally the average.
Execute this code twice: once with and once without monomorphization enabled and report the output of these two runs as well as the speedup.
function add(arg1, arg2) {
return arg1 + arg2;
}
var global = 0;
function action() {
for (var i = 0; i < 10000; i++) {
global = add(1, 2);
global = add("foo", "bar");
global = add([1,2,3], [4,5,6]);
}
}
// Warm up.
var start = Date.now();
while ((Date.now() - start) < 15000 /* 15 seconds */) {
action();
}
// Benchmark
var iterations = 0;
var sum = 0;
var start = Date.now();
while ((Date.now() - start) < 60000 /* 60 seconds */) {
var thisIterationStart = Date.now();
action();
var thisIterationTime = Date.now() - thisIterationStart;
iterations++;
sum += thisIterationTime;
}
console.log(sum / iterations);
The output without monomorphization is 4.494225288735564. The output with monomorphization is 4.2421633923. The speedup is ~5%.
Example 2 - Monomorphization of Indirect Calls #
This example is slightly more complicated and illustrates how monomorphization benefits higher order functions. In the example, the insertionSort
function is defined, which - given an array of items and a function for comparing these items - sorts the array using insertion sort.
Define an array of 1000 random double values between 0 and 1 and sort it four times using 4 different sorting methods (in the action
function).
Finally, as with the previous example, warm up the action
function for 15 second, and report the average execution time of
this function over the next 60 seconds with and without monomorphization.
function insertionSort (items, comparator) {
for (var i = 0; i < items.length; i++) {
let value = items[i];
for (var j = i - 1; j >= 0 && comparator(items[j], value); j--) {
items[j + 1] = items[j]
}
items[j + 1] = value
}
}
// Random values in an array
var array = new Array(1000);
for (i = 0; i < array.length; i++) {
array[i] = Math.random();
}
function action() {
insertionSort(array, function (a, b) { return a < b });
insertionSort(array, function (a, b) { return a > b });
insertionSort(array, function (a, b) { return a.toString().length < b.toString().length; });
insertionSort(array, function (a, b) { return a.toString().length > b.toString().length; });
}
// Warm up.
var start = Date.now();
while ((Date.now() - start) < 15000 /* 15 seconds */) {
action();
}
// Benchmark
var iterations = 0;
var sum = 0;
var start = Date.now();
while ((Date.now() - start) < 60000 /* 60 seconds */) {
var thisIterationStart = Date.now();
action();
var thisIterationTime = Date.now() - thisIterationStart;
iterations++;
sum += thisIterationTime;
}
console.log(sum / iterations);
The output without monomorphization is 194.05161290322582. The output with monomorphization is 175.41071428571428. The speedup is ~10%.